菜单
Run evaluations using the OpenHands/software-agent-sdk CI system. Supports SWE-bench, GAIA, Terminal-Bench, and other benchmarks via GitHub Actions.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/neubig/workflow --skill eval-with-ci命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Facilitate collaborative, human-in-the-loop Linear cycle or sprint planning. Use when the user asks to plan a cycle or sprint, choose cycle scope, set outcome themes, review capacity or carryover, sequence release work into a cycle, or work through a planning ticket surfaced by daily-workflow.
Human-in-the-loop daily workflow ("daily workflow", "my workflow", or "work queue") for checking review requests, aligning GitHub PRs and issues with Linear tickets, and triaging recent Slack requests before remediating the current user's eligible PRs.
Find Giant Eagle grocery products and ingredient substitutions for recipes using the website and the underlying product search API.
基于 SOC 职业分类
| name | eval-with-ci |
| description | Run evaluations using the OpenHands/software-agent-sdk CI system. Supports SWE-bench, GAIA, Terminal-Bench, and other benchmarks via GitHub Actions. |
| triggers | ["run eval","evaluation","benchmark","swebench","run-eval label"] |
Use OpenHands/software-agent-sdk GitHub Actions to run evaluations.
| Label | Instances | Use |
|---|---|---|
run-eval-1 | 1 | sanity check |
run-eval-50 | 50 | standard eval |
run-eval-200 | 200 | comprehensive eval |
run-eval-500 | 500 | full benchmark |
gh pr edit <PR_NUMBER> --repo OpenHands/software-agent-sdk --add-label run-eval-50
Actions → Run Eval → “Run workflow”, or CLI:
gh workflow run run-eval.yml \
--repo OpenHands/software-agent-sdk \
-f benchmark=swebench \
-f sdk_ref=v1.11.5 \
-f eval_limit=50 \
-f allow_unreleased_branches=false
Fields: benchmark (swebench, gaia, swtbench, commit0, multiswebench, terminalbench), sdk_ref (version or branch; branch requires allow_unreleased_branches=true), eval_limit, optional model_ids.
Supported: swebench (default SWE-bench Verified), gaia, swtbench, commit0, multiswebench, swebenchmultimodal, terminalbench.
Advanced fields: instance_ids (comma-separated), num_infer_workers, num_eval_workers, max_retries (default 3), tool_preset (default, gemini, gpt5, planning), eval_branch, benchmarks_branch.
gh run list --repo OpenHands/software-agent-sdk --workflow run-eval.yml --limit 5
gh run watch <RUN_ID> --repo OpenHands/software-agent-sdk
The PR gets a bot comment with trigger details; actual eval execution is dispatched to OpenHands/evaluation.
gh pr create --repo OpenHands/software-agent-sdk --title "My feature"
gh pr edit <PR_NUMBER> --repo OpenHands/software-agent-sdk --add-label run-eval-1
gh run list --repo OpenHands/software-agent-sdk --workflow run-eval.yml --limit 5
gh run watch <RUN_ID> --repo OpenHands/software-agent-sdk
| Issue | Fix |
|---|---|
| Unsupported label | Use exact run-eval-1/50/200/500. |
| Model ID not found | Check .github/run-eval/resolve_model_config.py. |
| SDK reference invalid | Use v1.x.x or allow unreleased branches. |
| Workflow does not start | Confirm repo write access / workflow-scoped token. |